Clément Godard

Because it is essentially python, it is very easy to use and you can iterate and try things extremely quickly. The dataloaders are also very nice as they are pure...

@abhiagwl4262 because we are using tf.abs you can simply compute the mean separately for both or you can also crop each of them so that the sizes matches (H-1, W-1)

@mojovski apologies for the massive delay. Essentially the behaviour is undefined with images which have different aspect ratios and focals as the network really only saw one type of images....

I am not sure I follow what you meant by re calibrating. I think the most general case would require to train on a variety of aspect ratios and focals...

Hi, Did you have a look at the generated depthmaps?

Hi @7LFB, I think you are correct. My implementation of Resnet50 was quite last minute and has actually another bug. I will leave it as is for now to make...

Hi, Yes it would definitely help, you would need to load the ground truth depths when training and adding a regression loss on the non-zero depth points. Have a look...

I am using pythorch 1.7.1 as my GPU only supports CUDA 11.0 and up. However I have used newer versions of pytorch on older ones and never had issues. Are...

Thanks for your help! I see that [someone seems to have a similar issue for NAS](https://github.com/udaykusupati/Normal-Assisted-Stereo/issues/16) on pytorch 1.3.1 so it's possible the backwards compatibility breaking change happened between 1.1....

This looks a lot better, thanks!